JDK 9 RFR of JDK-8074977: Constructor.getAnnotatedParameterTypes returns wrong value

joe darcy joe.darcy at oracle.com
Sun May 21 00:23:28 UTC 2017


Hi Peter,

Yours is a more ambitious attempt to more fully resolve this issue. It 
would likely nearly always work in practice, at least for 
javac-generated class files.

However, at this stage of JDK 9, I'd be more comfortable with something 
like the point fixes in my current patch, deferring a more complete 
solution to JDK 10. At that point, I think a two-stage fall back would 
be more complete: if the number of parameters differs, first try to see 
if java.lang.reflect.Parameter.{isSynthetic, isImplicit} information is 
available and if it is, use it accordingly. Otherwise, use the sort of 
case analysis present in the patch.

FWIW, on the javax.lang.model side of reflective APIs, in JDK 9 we added 
notions in the API to describe these sorts of differences in the 
javax.lang.model.util.Elements.Origin enum:

http://download.java.net/java/jdk9/docs/api/javax/lang/model/util/Elements.Origin.html

and the Elements.getOrigin methods which return that type.

Thanks,

-Joe


On 5/20/2017 2:48 PM, Peter Levart wrote:
> Hi Joe,
>
> Thanks for explanation.
>
> Regarding your patch. Take a look here:
>
> http://cr.openjdk.java.net/~plevart/jdk9-dev/8074977_ConstructorParameterAnnotations/webrev.01/
>
> I just want to point out that the logic could be centralized for 
> classical and type annotations. What do you think?
>
>
> Regards, Peter
>
>
> On 05/20/2017 11:38 PM, joe darcy wrote:
>> Hi Peter,
>>
>>
>> On 5/20/2017 11:00 AM, Peter Levart wrote:
>>> Hi Joe,
>>>
>>> So enum classes, anonymous classes and local classes don't have a 
>>> standard way of passing synthetic/implicit constructor parameters?
>>
>> Right; there is no standardized or mandated way for compilers to do 
>> that. Such details are intentionally outside of what the JLS specifies.
>>
>> Since non-static member classes can be referred to and instances 
>> instantiated from outside of the enclosing class, there has to be a 
>> standardized linkage model in that case for separate compilation, 
>> etc. and that is covered in the JLS. This isn't the case for local 
>> and anonymous classes since there isn't a way to refer to such 
>> classes from outside of their immediate declaration context. The 
>> instances of such types can be returned of course and then it is 
>> possible for people to getClass().getConstructors()... their way into 
>> the issue.
>>
>>
>>> Do various compilers do it differently than javac?
>>
>> IIRC, ecj does compile enum classes differently than javac.
>>
>>>
>>> It's unfortunate but if # of parameter annotations arrays is less 
>>> than the # of parameters, Parameter.getAnnotations() may return 
>>> wrong annotations or throw IndexOutOfBoundException for enums, 
>>> anonymous and local classes. Can't we do anything for them? At least 
>>> for code compiled by javac?
>>>
>>> For example, javac compiles enum classes so that they always prefix 
>>> constructor source parameters with a pair of (String name, int 
>>> ordinal, ...) and so do anonymous enum subclasses provided by enum 
>>> constants (i.e. clazz.isAnonymousClass() && 
>>> clazz.getSuperclass().isEnum())
>>>
>>> Non-static local classes are compiled so that constructor source 
>>> parameters are prefixed with a single parameter (OuterClass 
>>> outerInstance, ...), while any captured variables follow source 
>>> parameters
>>
>> A wrinkle with local classes is that if they occur in a static 
>> context, such as a static initializer block, there is no leading 
>> outer instance parameter.
>>
>> This is admittedly not a 100% solution to parameter mismatches and 
>> annotations; however, it is an improvement over the current state of 
>> affairs.
>>
>> Ideally, the code for something like Parameters.getAnnotations could 
>> do something like look at the synthetic bit and pair up natural 
>> parameters with annotations if there was a mismatch in the number of 
>> parameters and annotations. Unfortunately, the Parameters.isSynthetic 
>> API wasn't introduced until JDK 8 and that information doesn't have 
>> to be available.
>>
>> Cheers,
>>
>> -Joe
>>
>>>
>>> Static local classes are compiled so that constructor source 
>>> parameters come 1st, followed by any captured variables.
>>>
>>> etc...
>>>
>>> Regards, Peter
>>>
>>>
>>> On 05/19/2017 11:31 PM, joe darcy wrote:
>>>> Hello,
>>>>
>>>> Please review the webrev to fix
>>>>
>>>>     JDK-8074977: Constructor.getAnnotatedParameterTypes returns 
>>>> wrong value
>>>> http://cr.openjdk.java.net/~darcy/8074977.3/
>>>>
>>>> To condense a complicated situation, there are cases where the 
>>>> number of parameters present for a constructor in source code and 
>>>> the number of parameters present for that constructor in the class 
>>>> file differ. One of those cases occurs for the constructor of a 
>>>> non-static member class [1] where there is a leading parameter to 
>>>> accept the outer this argument.
>>>>
>>>> Bug JDK-8074977 reports on a situation where the type annotations 
>>>> on constructor parameters are incorrectly reported. Essentially, an 
>>>> off-by-one error is the cause since the annotation information is 
>>>> stored with respect to the number of parameters present in the 
>>>> source and an additional parameter is present at runtime.
>>>>
>>>> An analogous situation exists for declaration annotations and 
>>>> constructor parameters, declaration annotations being the 
>>>> traditional flavor of annotations.
>>>>
>>>> Type annotations and declaration annotations are read using 
>>>> different APIs so require separate fixes to detect the additional 
>>>> parameter and make the necessary adjustments in the returned 
>>>> information.
>>>>
>>>> The regression tests cover both the type annotation reading API and 
>>>> the two ways of reading declaration annotations on parameters, 
>>>> calling getParameterAnnotations on the constructor or call 
>>>> getParameters on the constructor and then calling getAnnotations on 
>>>> each parameter. The getParameters API was added in JDK 8.
>>>>
>>>> Static member and non-static member classes are used as test cases, 
>>>> as are constructors with and without generic type information.
>>>>
>>>> Thanks,
>>>>
>>>> -Joe
>>>>
>>>> [1] 
>>>> https://blogs.oracle.com/darcy/nested,-inner,-member,-and-top-level-classes
>>>>
>>>
>>
>



More information about the core-libs-dev mailing list